Computer Vision Applications in Industry
Where computer vision pays off in manufacturing, retail, healthcare and vehicles, how to scope a project so it can succeed, and a short people-counting prototype to test an idea in an afternoon.

A retail contact once asked me whether a camera could tell them how many people join the checkout queue each hour. The first answer I gave was "probably". The honest answer took a few days to reach: yes for counting, no for guessing who is a customer versus staff, and only if the camera angle, lighting and privacy rules were settled first. The model was the easy part. The hard part was agreeing on exactly what question the camera had to answer.
This post surveys where computer vision works in industry, then shows how I scope a project and build a first prototype before anyone commits to a data collection campaign. For deep dives on factory inspection, see Building Resilient Computer Vision Pipelines for High-Speed Manufacturing and Rare Defect Detection With Few Labels.
Quick takeaways
- Computer vision succeeds when the question is narrow and measurable ("count people crossing this line"), not when it is vague ("understand the store").
- Data matters more than the model: representative images, varied conditions and consistent labels decide the outcome.
- Decide where inference runs (edge or cloud) early; it drives latency, cost and privacy.
- Prototype with a pretrained detector and a fixed success metric before collecting any custom data.
- Regulated domains such as healthcare need validation and compliance work that dwarfs the modelling effort.
Which industries use computer vision, and for what?
The patterns repeat across sectors. Most projects reduce to one of four tasks: classify an image, detect and locate objects, segment pixels, or track objects over time.
| Domain | Typical application | Main task | Usual difficulty |
|---|---|---|---|
| Manufacturing | Surface defects, assembly verification, dimension checks | Detection, segmentation, anomaly detection | Rare defects, lighting control |
| Retail | Foot traffic, queue length, shelf stock, planogram compliance | Detection, tracking | Occlusion, privacy |
| Healthcare | X-ray triage, pathology slides, retinal screening | Classification, segmentation | Regulation, interpretability, data access |
| Vehicles and logistics | Lane and pedestrian detection, parcel scanning, damage inspection | Detection, tracking, OCR | Edge cases, real-time limits |
| Agriculture and infrastructure | Crop health, crack detection, drone inspection | Segmentation, classification | Weather, domain shift |
Manufacturing is where I have seen the clearest return because the environment is controlled: fixed cameras, repeatable parts, and a defect cost that is easy to compute. Healthcare has the highest stakes. A model there is a medical device in the eyes of the regulator, so I treat it as a compliance project with a model inside, not the other way around.
How do I scope a computer vision project so it can succeed?
Write the problem down before touching code. I use a one-page spec like this, and I refuse to start until every line has an answer:
question: How many people join the checkout queue per hour?
output: integer count per 15-minute window
success_metric: absolute error under 10% against a manual count over 3 days
camera: fixed, 1080p, mounted above the queue, 25 fps
conditions: daylight and artificial light, up to 15 people in frame
privacy: no frames stored; only counts leave the device
runs_on: edge device at the store
failure_cost: a wrong count misleads staffing, but nobody is harmed
Two lines in that file save the most pain. The success metric must be measured against something real, not against the model's own output. And the privacy line decides whether you process on the device or send video to a server. When you can answer "what happens when the model is wrong", you can decide how much accuracy you actually need.
Edge or cloud: where should inference run?
Edge inference gives low latency, works offline and keeps raw video local, which is also the easiest privacy story. Cloud inference gives more compute and makes updates trivial, at the cost of bandwidth and a round trip. A common compromise is to detect on the device and send only metadata (counts, crops of events) to the cloud. If you need to squeeze a model onto small hardware, ONNX Runtime and TensorRT covers exporting and quantizing.
What does a quick prototype look like?
Before collecting custom data, test whether a pretrained model already solves most of the problem. This script counts people crossing a horizontal line in a video using Ultralytics tracking:
pip install ultralytics
from ultralytics import YOLO
model = YOLO("yolo11n.pt") # small pretrained detector; swap for a current weights file
LINE_Y = 400 # pixel row of the counting line in your frame
last_y: dict[int, float] = {}
count = 0
for result in model.track(source="queue.mp4", stream=True, persist=True, classes=[0], verbose=False):
boxes = result.boxes
if boxes.id is None:
continue
for track_id, (cx, cy, w, h) in zip(boxes.id.int().tolist(), boxes.xywh.tolist()):
previous = last_y.get(track_id)
if previous is not None and previous < LINE_Y <= cy:
count += 1 # moved downwards across the line
last_y[track_id] = cy
print(f"people crossing the line: {count}")
Class 0 is "person" in the COCO classes this model uses, and persist=True keeps tracker identities between frames so a person is counted once. Compare the number with a manual count on a few minutes of footage. If it is close, you have a case for a pilot. If it fails, look at why before blaming the model: usually the camera angle, people overlapping, or the line placed where tracks flicker. The object detection basics behind this are explained in YOLO Object Detection Guide.
What should I do this week?
Take the project idea you have and fill in the one-page spec above. Then record five minutes of real footage from the planned camera position and run the prototype on it. By the end of the week you will know whether you need a bigger model, better data, a different camera angle, or a different project, which is the cheapest moment to learn it.
Uğur Kaval
AI/ML Engineer & Full Stack Developer specializing in building innovative solutions with modern technologies. Passionate about automation, machine learning, and web development.


